Missing data is a common challenge in time series collected from real-world sensing systems, especially in environmental and agricultural monitoring. This study proposes a novel framework for evaluating the reliability of time series imputation methods in the absence of ground truth data. The framework consists of two stages: (1) imputing missing values using a diverse set of machine learning (ML) and deep learning (DL) models, including Support Vector Regression (SVR), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), one-dimensional Convolutional Neural Network (1DCNN), Gated Recurrent Unit (GRU), Attention, and Multi-head Attention (MA), based on the weighted bi-directional imputation (WBDI) method; and (2) assessing the imputation data quality through statistical distribution analysis and a regression task: predicting volumetric water content. Experiments are conducted on the CAF003 dataset, a long-term (2007–2016) soil sensor time series from the Cook Agronomy Farm (WSU, USA), which exhibits 41–47% missing data across multiple depths. The results indicate that Attention-based imputation method consistently achieves higher prediction accuracy while maintaining the statistical characteristics of the original data. This consistent performance across multiple evaluation metrics highlights its effectiveness as a robust method for recovering missing values in smart environmental sensing systems.